Role Overview The Office of Responsible AI within Digital Foundation Services is looking for a practitioner who can hold both halves of the job at once: the engineering depth to take an agentic AI system apart and see how it actually behaves, and the governance fluency to judge whether it should be allowed to run. This is not a policy-only role, and it is not a pure engineering role. The person we are looking for reviews architectures, reads evaluation results, challenges design choices, and then turns those findings into control decisions that clients, delivery teams and regulators can rely on.

The role works across delivery teams shipping agentic and generative AI into production for enterprise customers. Day to day the work spans risk assessment of new AI use cases, hands-on review of model and retrieval pipelines, definition of guardrails and evaluation criteria, and enablement of engineering teams so that responsible AI practice is built in rather than inspected in at the end.

Key Responsibilities

Governance and assurance

- Run intake and risk classification for new AI and agentic use cases, applying an agreed risk-tiering method and recording the rationale, so decisions remain auditable.
- Own the responsible AI control framework — policies, standards, checklists and approval gates — and keep it aligned to EU AI Act, NIST AI RMF and ISO/IEC 42001 expectations.
- Chair or support the AI review board: prepare evidence packs, present findings, and track conditions of approval through closure.
- Maintain the AI inventory and model registry, covering ownership, data lineage, intended use and residual risk for every deployed system.
- Produce audit-ready documentation: model cards, system cards, data sheets, impact assessments, and evaluation reports.

Technical assurance of AI systems
- Review agentic architectures end to end: tool and function calling, planning loops, memory, autonomy boundaries, human-in-the-loop checkpoints and failure containment.
- Assess retrieval-augmented generation pipelines for grounding quality, chunking and embedding choices, retrieval relevance, citation fidelity and leakage of restricted content.
- Evaluate fine-tuned and adapted models for fitness of purpose, drift, catastrophic forgetting, and the provenance and licensing of training data.
- Design and oversee evaluation of harnesses covering accuracy, robustness, bias, toxicity, prompt injection, jailbreak resistance and data exfiltration.
- Define guardrail patterns — input and output filtering, policy engines, tool sandboxing, scope and rate limits — and validate that they hold under adversarial testing.

Enablement and advisory
- Advise solution architects and delivery lead at design time, so controls are chosen early rather than retrofitted late.
- Build and run enablement material: playbooks, reference patterns, training and office hours for the engineering and delivery communities.
- Act as escalation point for AI incidents, running root-cause analysis and feeding lessons back into the control framework.
- Support client conversations and bid responses where responsible AI posture is a differentiator or a contractual requirement.

Technical Skills Required Agentic AI frameworks and tooling
- Working knowledge of at least two agentic frameworks — LangGraph, LangChain, AutoGen, CrewAI, Semantic Kernel or the OpenAI Agents SDK — including their orchestration and state models.
- Familiarity with agent interoperability and tool-integration standards,



including Model Context Protocol and comparable function-calling interfaces.
- Practical experience with multi-agent patterns such as planner-executor, supervisor-worker and reflection loops, and with the failure modes each introduces.
- Ability to read production agent code and execution traces, and to instrument systems with observability tooling such as LangSmith, Langfuse or OpenTelemetry.

Models, fine-tuning and RAG
- Solid grounding in transformer architectures, tokenization, context windows, decoding parameters and the practical limits of current frontier and open-weight models.
- Hands-on understanding of adaptation techniques — prompt engineering, instruction tuning, LoRA and QLoRA, full fine-tuning, RLHF and preference optimization — and when each is appropriate.
- Experience building or reviewing RAG systems: embedding models, vector stores such as FAISS, Azure AI Search, Pinecone or pgvector, hybrid and re-ranked retrieval, and evaluation of grounding quality.
- Comfort with evaluation tooling and benchmarks, including LLM-as-judge techniques and a clear view of their known weaknesses.
- Awareness of cost, latency and sustainability trade-offs across hosted and self-hosted deployment options.

Engineering foundations
- Python at a level sufficient to prototype, read production code and build evaluation scripts.
- Familiarity with at least one major cloud AI platform — Azure AI Foundry, AWS Bedrock or Google Vertex AI — and its native governance and content-safety controls.
- Understanding of MLOps and LLMOps practice: CI/CD, versioning, monitoring, rollback and change control for models and prompts.
- Working knowledge of security and privacy fundamentals as they apply to AI: identity, secrets, data classification, anonymization and residency.

AI Governance and Responsible AI Expertise
- Practical command of the EU AI Act, including risk tiering, general-purpose AI obligations and conformity expectations, with the ability to translate them into engineering controls.
- Familiarity with NIST AI RMF, ISO/IEC 42001 and ISO/IEC 23894, and experience operationalizing at least one of them.
- Understanding of sector regulation relevant to financial services and other regulated industries, including model risk management expectations such as SR 11-7.
- Knowledge of data protection regimes such as GDPR and the DPDP Act, and how they interact with training data, inference logs and cross-border transfer.
- Fluency in fairness, explainability, transparency and accountability techniques, with the judgement to select the right one rather than defaulting to a favorite.
- Experience running AI impact assessments and third-party or vendor AI risk reviews.
- Awareness of emerging assurance practice: red teaming, model evaluation standards, content provenance and watermarking.

Soft Skills and Behavioural Competencies This role only works if the person can carry a technical argument into a business room and be believed in both. The behaviours below carry as much weight in selection as the technical depth above.
- Influence without authority.



Much of the job is persuading teams that do not report to you to change a design, which takes credibility, patience and a workable alternative rather than a veto.
- Clear communication across levels. The same finding must be expressed as an engineering ticket, a client-facing paragraph and a board-level risk statement without losing accuracy at any step.
- Constructive challenge. Comfortable saying no to a senior stakeholder, and equally comfortable saying yes quickly when the risk is genuinely low; proportionality matters more than caution.
- Judgement under ambiguity. Regulation, tooling and model behaviour are all moving. Strong candidates make defensible decisions on incomplete information and revisit them as evidence changes.
- Intellectual honesty. Willing to state the limits of an evaluation, flag their own uncertainty, and correct a published position when new evidence arrives.
- Cross-functional collaboration. Works naturally with legal, privacy, security, procurement, HR and delivery, and can translate between their vocabularies.
- Ethical grounding and empathy. Genuinely considers the people affected by an AI system, including those who are not the paying customer.
- Pragmatism and delivery focus. Understands commercial pressure and designs controls teams can actually adopt, rather than an ideal framework nobody follows.
- Facilitation and coaching. Runs review boards, workshops and enablement sessions that leave people clearer than they arrived.
- Continuous learning. Keeps current with research, incidents and regulatory movement, and shares that knowledge rather than hoarding it.

Qualifications and Experience
- Bachelor's or Master's degree in computer science, data science, engineering or a related quantitative discipline; a law, policy or ethics background combined with demonstrable technical depth will also be considered.
- Eight to twelve years of overall technology experience, including at least three years working directly on machine learning or generative AI systems.
- At least two years in an AI governance, model risk, AI assurance or responsible AI capacity, or equivalent evidence of leading such work from inside an engineering team.
- Demonstrable experience taking at least one generative or agentic AI system from design through to governed production deployment.
- Client-facing or consulting experience is an advantage, given the role's exposure to customer, audit and regulator conversations.

Preferred Certifications and Credentials
- IAPP AI Governance Qualified (AIGP).
- ISO/IEC 42001 lead implementer or lead auditor.
- A cloud AI engineering certification such as Azure AI Engineer Associate, AWS Machine Learning Specialty or Google Professional Machine Learning Engineer.
- Security or privacy credentials such as CISSP, CIPP/E or CIPM.
- Published work, open-source contributions or conference talks on responsible AI, model evaluation or agent safety.

What Success Looks Like in the First Year
- Every in-flight AI and agentic use case is classified, inventoried and mapped to a control set with a named owner.
- A documented, repeatable review gate exists and is used by delivery teams without escalation in the majority of cases.
- Evaluation and red-teaming harnesses run automatically against production agent workloads, with results visible to engineering leadership.
- Delivery teams arrive at the review board with evidence already prepared, because the enablement material made the expectations obvious.

📌 Responsible AI & Governance (Noida)
🏢 HCLTech
📍 Noida